backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32 #include <pthread.h>
  33
  34 #include "fio.h"
  35 #include "smalloc.h"
  36 #include "verify.h"
  37 #include "diskutil.h"
  38 #include "cgroup.h"
  39 #include "profile.h"
  40 #include "lib/rand.h"
  41 #include "lib/memalign.h"
  42 #include "server.h"
  43 #include "lib/getrusage.h"
  44 #include "idletime.h"
  45 #include "err.h"
  46 #include "workqueue.h"
  47 #include "lib/mountcheck.h"
  48 #include "rate-submit.h"
  49 #include "helper_thread.h"
  50 #include "pshared.h"
  51 #include "zone-dist.h"
  52
  53 static struct fio_sem *startup_sem;
  54 static struct flist_head *cgroup_list;
  55 static struct cgroup_mnt *cgroup_mnt;
  56 static int exit_value;
  57 static volatile bool fio_abort;
  58 static unsigned int nr_process = 0;
  59 static unsigned int nr_thread = 0;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int stat_number = 0;
  66 int shm_id = 0;
  67 int temp_stall_ts;
  68 unsigned long done_secs = 0;
  69 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
  70 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
  71 #else
  72 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  73 #endif
  74
  75 #define JOB_START_TIMEOUT       (5 * 1000)
  76
  77 static void sig_int(int sig)
  78 {
  79         if (threads) {
  80                 if (is_backend)
  81                         fio_server_got_signal(sig);
  82                 else {
  83                         log_info("\nfio: terminating on signal %d\n", sig);
  84                         log_info_flush();
  85                         exit_value = 128;
  86                 }
  87
  88                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  89         }
  90 }
  91
  92 void sig_show_status(int sig)
  93 {
  94         show_running_run_stats();
  95 }
  96
  97 static void set_sig_handlers(void)
  98 {
  99         struct sigaction act;
 100
 101         memset(&act, 0, sizeof(act));
 102         act.sa_handler = sig_int;
 103         act.sa_flags = SA_RESTART;
 104         sigaction(SIGINT, &act, NULL);
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGTERM, &act, NULL);
 110
 111 /* Windows uses SIGBREAK as a quit signal from other applications */
 112 #ifdef WIN32
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGBREAK, &act, NULL);
 117 #endif
 118
 119         memset(&act, 0, sizeof(act));
 120         act.sa_handler = sig_show_status;
 121         act.sa_flags = SA_RESTART;
 122         sigaction(SIGUSR1, &act, NULL);
 123
 124         if (is_backend) {
 125                 memset(&act, 0, sizeof(act));
 126                 act.sa_handler = sig_int;
 127                 act.sa_flags = SA_RESTART;
 128                 sigaction(SIGPIPE, &act, NULL);
 129         }
 130 }
 131
 132 /*
 133  * Check if we are above the minimum rate given.
 134  */
 135 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 136                              enum fio_ddir ddir)
 137 {
 138         unsigned long long bytes = 0;
 139         unsigned long iops = 0;
 140         unsigned long spent;
 141         unsigned long long rate;
 142         unsigned long long ratemin = 0;
 143         unsigned int rate_iops = 0;
 144         unsigned int rate_iops_min = 0;
 145
 146         assert(ddir_rw(ddir));
 147
 148         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 149                 return false;
 150
 151         /*
 152          * allow a 2 second settle period in the beginning
 153          */
 154         if (mtime_since(&td->start, now) < 2000)
 155                 return false;
 156
 157         iops += td->this_io_blocks[ddir];
 158         bytes += td->this_io_bytes[ddir];
 159         ratemin += td->o.ratemin[ddir];
 160         rate_iops += td->o.rate_iops[ddir];
 161         rate_iops_min += td->o.rate_iops_min[ddir];
 162
 163         /*
 164          * if rate blocks is set, sample is running
 165          */
 166         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 167                 spent = mtime_since(&td->lastrate[ddir], now);
 168                 if (spent < td->o.ratecycle)
 169                         return false;
 170
 171                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 172                         /*
 173                          * check bandwidth specified rate
 174                          */
 175                         if (bytes < td->rate_bytes[ddir]) {
 176                                 log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n",
 177                                         td->o.name, ratemin, bytes);
 178                                 return true;
 179                         } else {
 180                                 if (spent)
 181                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 182                                 else
 183                                         rate = 0;
 184
 185                                 if (rate < ratemin ||
 186                                     bytes < td->rate_bytes[ddir]) {
 187                                         log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 188                                                 td->o.name, ratemin, rate);
 189                                         return true;
 190                                 }
 191                         }
 192                 } else {
 193                         /*
 194                          * checks iops specified rate
 195                          */
 196                         if (iops < rate_iops) {
 197                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 198                                                 td->o.name, rate_iops, iops);
 199                                 return true;
 200                         } else {
 201                                 if (spent)
 202                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 203                                 else
 204                                         rate = 0;
 205
 206                                 if (rate < rate_iops_min ||
 207                                     iops < td->rate_blocks[ddir]) {
 208                                         log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 209                                                 td->o.name, rate_iops_min, rate);
 210                                         return true;
 211                                 }
 212                         }
 213                 }
 214         }
 215
 216         td->rate_bytes[ddir] = bytes;
 217         td->rate_blocks[ddir] = iops;
 218         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 219         return false;
 220 }
 221
 222 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 223 {
 224         bool ret = false;
 225
 226         if (td->bytes_done[DDIR_READ])
 227                 ret |= __check_min_rate(td, now, DDIR_READ);
 228         if (td->bytes_done[DDIR_WRITE])
 229                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 230         if (td->bytes_done[DDIR_TRIM])
 231                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 232
 233         return ret;
 234 }
 235
 236 /*
 237  * When job exits, we can cancel the in-flight IO if we are using async
 238  * io. Attempt to do so.
 239  */
 240 static void cleanup_pending_aio(struct thread_data *td)
 241 {
 242         int r;
 243
 244         /*
 245          * get immediately available events, if any
 246          */
 247         r = io_u_queued_complete(td, 0);
 248
 249         /*
 250          * now cancel remaining active events
 251          */
 252         if (td->io_ops->cancel) {
 253                 struct io_u *io_u;
 254                 int i;
 255
 256                 io_u_qiter(&td->io_u_all, io_u, i) {
 257                         if (io_u->flags & IO_U_F_FLIGHT) {
 258                                 r = td->io_ops->cancel(td, io_u);
 259                                 if (!r)
 260                                         put_io_u(td, io_u);
 261                         }
 262                 }
 263         }
 264
 265         if (td->cur_depth)
 266                 r = io_u_queued_complete(td, td->cur_depth);
 267 }
 268
 269 /*
 270  * Helper to handle the final sync of a file. Works just like the normal
 271  * io path, just does everything sync.
 272  */
 273 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 274 {
 275         struct io_u *io_u = __get_io_u(td);
 276         enum fio_q_status ret;
 277
 278         if (!io_u)
 279                 return true;
 280
 281         io_u->ddir = DDIR_SYNC;
 282         io_u->file = f;
 283         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 284
 285         if (td_io_prep(td, io_u)) {
 286                 put_io_u(td, io_u);
 287                 return true;
 288         }
 289
 290 requeue:
 291         ret = td_io_queue(td, io_u);
 292         switch (ret) {
 293         case FIO_Q_QUEUED:
 294                 td_io_commit(td);
 295                 if (io_u_queued_complete(td, 1) < 0)
 296                         return true;
 297                 break;
 298         case FIO_Q_COMPLETED:
 299                 if (io_u->error) {
 300                         td_verror(td, io_u->error, "td_io_queue");
 301                         return true;
 302                 }
 303
 304                 if (io_u_sync_complete(td, io_u) < 0)
 305                         return true;
 306                 break;
 307         case FIO_Q_BUSY:
 308                 td_io_commit(td);
 309                 goto requeue;
 310         }
 311
 312         return false;
 313 }
 314
 315 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 316 {
 317         int ret, ret2;
 318
 319         if (fio_file_open(f))
 320                 return fio_io_sync(td, f);
 321
 322         if (td_io_open_file(td, f))
 323                 return 1;
 324
 325         ret = fio_io_sync(td, f);
 326         ret2 = 0;
 327         if (fio_file_open(f))
 328                 ret2 = td_io_close_file(td, f);
 329         return (ret || ret2);
 330 }
 331
 332 static inline void __update_ts_cache(struct thread_data *td)
 333 {
 334         fio_gettime(&td->ts_cache, NULL);
 335 }
 336
 337 static inline void update_ts_cache(struct thread_data *td)
 338 {
 339         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 340                 __update_ts_cache(td);
 341 }
 342
 343 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 344 {
 345         if (in_ramp_time(td))
 346                 return false;
 347         if (!td->o.timeout)
 348                 return false;
 349         if (utime_since(&td->epoch, t) >= td->o.timeout)
 350                 return true;
 351
 352         return false;
 353 }
 354
 355 /*
 356  * We need to update the runtime consistently in ms, but keep a running
 357  * tally of the current elapsed time in microseconds for sub millisecond
 358  * updates.
 359  */
 360 static inline void update_runtime(struct thread_data *td,
 361                                   unsigned long long *elapsed_us,
 362                                   const enum fio_ddir ddir)
 363 {
 364         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 365                 return;
 366
 367         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 368         elapsed_us[ddir] += utime_since_now(&td->start);
 369         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 370 }
 371
 372 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 373                                 int *retptr)
 374 {
 375         int ret = *retptr;
 376
 377         if (ret < 0 || td->error) {
 378                 int err = td->error;
 379                 enum error_type_bit eb;
 380
 381                 if (ret < 0)
 382                         err = -ret;
 383
 384                 eb = td_error_type(ddir, err);
 385                 if (!(td->o.continue_on_error & (1 << eb)))
 386                         return true;
 387
 388                 if (td_non_fatal_error(td, eb, err)) {
 389                         /*
 390                          * Continue with the I/Os in case of
 391                          * a non fatal error.
 392                          */
 393                         update_error_count(td, err);
 394                         td_clear_error(td);
 395                         *retptr = 0;
 396                         return false;
 397                 } else if (td->o.fill_device && err == ENOSPC) {
 398                         /*
 399                          * We expect to hit this error if
 400                          * fill_device option is set.
 401                          */
 402                         td_clear_error(td);
 403                         fio_mark_td_terminate(td);
 404                         return true;
 405                 } else {
 406                         /*
 407                          * Stop the I/O in case of a fatal
 408                          * error.
 409                          */
 410                         update_error_count(td, err);
 411                         return true;
 412                 }
 413         }
 414
 415         return false;
 416 }
 417
 418 static void check_update_rusage(struct thread_data *td)
 419 {
 420         if (td->update_rusage) {
 421                 td->update_rusage = 0;
 422                 update_rusage_stat(td);
 423                 fio_sem_up(td->rusage_sem);
 424         }
 425 }
 426
 427 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 428 {
 429         const int full = queue_full(td);
 430         int min_evts = 0;
 431         int ret;
 432
 433         if (td->flags & TD_F_REGROW_LOGS)
 434                 return io_u_quiesce(td);
 435
 436         /*
 437          * if the queue is full, we MUST reap at least 1 event
 438          */
 439         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 440         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 441                 min_evts = 1;
 442
 443         if (time && __should_check_rate(td))
 444                 fio_gettime(time, NULL);
 445
 446         do {
 447                 ret = io_u_queued_complete(td, min_evts);
 448                 if (ret < 0)
 449                         break;
 450         } while (full && (td->cur_depth > td->o.iodepth_low));
 451
 452         return ret;
 453 }
 454
 455 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 456                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 457                    struct timespec *comp_time)
 458 {
 459         switch (*ret) {
 460         case FIO_Q_COMPLETED:
 461                 if (io_u->error) {
 462                         *ret = -io_u->error;
 463                         clear_io_u(td, io_u);
 464                 } else if (io_u->resid) {
 465                         long long bytes = io_u->xfer_buflen - io_u->resid;
 466                         struct fio_file *f = io_u->file;
 467
 468                         if (bytes_issued)
 469                                 *bytes_issued += bytes;
 470
 471                         if (!from_verify)
 472                                 trim_io_piece(io_u);
 473
 474                         /*
 475                          * zero read, fail
 476                          */
 477                         if (!bytes) {
 478                                 if (!from_verify)
 479                                         unlog_io_piece(td, io_u);
 480                                 td_verror(td, EIO, "full resid");
 481                                 put_io_u(td, io_u);
 482                                 break;
 483                         }
 484
 485                         io_u->xfer_buflen = io_u->resid;
 486                         io_u->xfer_buf += bytes;
 487                         io_u->offset += bytes;
 488
 489                         if (ddir_rw(io_u->ddir))
 490                                 td->ts.short_io_u[io_u->ddir]++;
 491
 492                         if (io_u->offset == f->real_file_size)
 493                                 goto sync_done;
 494
 495                         requeue_io_u(td, &io_u);
 496                 } else {
 497 sync_done:
 498                         if (comp_time && __should_check_rate(td))
 499                                 fio_gettime(comp_time, NULL);
 500
 501                         *ret = io_u_sync_complete(td, io_u);
 502                         if (*ret < 0)
 503                                 break;
 504                 }
 505
 506                 if (td->flags & TD_F_REGROW_LOGS)
 507                         regrow_logs(td);
 508
 509                 /*
 510                  * when doing I/O (not when verifying),
 511                  * check for any errors that are to be ignored
 512                  */
 513                 if (!from_verify)
 514                         break;
 515
 516                 return 0;
 517         case FIO_Q_QUEUED:
 518                 /*
 519                  * if the engine doesn't have a commit hook,
 520                  * the io_u is really queued. if it does have such
 521                  * a hook, it has to call io_u_queued() itself.
 522                  */
 523                 if (td->io_ops->commit == NULL)
 524                         io_u_queued(td, io_u);
 525                 if (bytes_issued)
 526                         *bytes_issued += io_u->xfer_buflen;
 527                 break;
 528         case FIO_Q_BUSY:
 529                 if (!from_verify)
 530                         unlog_io_piece(td, io_u);
 531                 requeue_io_u(td, &io_u);
 532                 td_io_commit(td);
 533                 break;
 534         default:
 535                 assert(*ret < 0);
 536                 td_verror(td, -(*ret), "td_io_queue");
 537                 break;
 538         }
 539
 540         if (break_on_this_error(td, ddir, ret))
 541                 return 1;
 542
 543         return 0;
 544 }
 545
 546 static inline bool io_in_polling(struct thread_data *td)
 547 {
 548         return !td->o.iodepth_batch_complete_min &&
 549                    !td->o.iodepth_batch_complete_max;
 550 }
 551 /*
 552  * Unlinks files from thread data fio_file structure
 553  */
 554 static int unlink_all_files(struct thread_data *td)
 555 {
 556         struct fio_file *f;
 557         unsigned int i;
 558         int ret = 0;
 559
 560         for_each_file(td, f, i) {
 561                 if (f->filetype != FIO_TYPE_FILE)
 562                         continue;
 563                 ret = td_io_unlink_file(td, f);
 564                 if (ret)
 565                         break;
 566         }
 567
 568         if (ret)
 569                 td_verror(td, ret, "unlink_all_files");
 570
 571         return ret;
 572 }
 573
 574 /*
 575  * Check if io_u will overlap an in-flight IO in the queue
 576  */
 577 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 578 {
 579         bool overlap;
 580         struct io_u *check_io_u;
 581         unsigned long long x1, x2, y1, y2;
 582         int i;
 583
 584         x1 = io_u->offset;
 585         x2 = io_u->offset + io_u->buflen;
 586         overlap = false;
 587         io_u_qiter(q, check_io_u, i) {
 588                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 589                         y1 = check_io_u->offset;
 590                         y2 = check_io_u->offset + check_io_u->buflen;
 591
 592                         if (x1 < y2 && y1 < x2) {
 593                                 overlap = true;
 594                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 595                                                 x1, io_u->buflen,
 596                                                 y1, check_io_u->buflen);
 597                                 break;
 598                         }
 599                 }
 600         }
 601
 602         return overlap;
 603 }
 604
 605 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 606 {
 607         /*
 608          * Check for overlap if the user asked us to, and we have
 609          * at least one IO in flight besides this one.
 610          */
 611         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 612             in_flight_overlap(&td->io_u_all, io_u))
 613                 return FIO_Q_BUSY;
 614
 615         return td_io_queue(td, io_u);
 616 }
 617
 618 /*
 619  * The main verify engine. Runs over the writes we previously submitted,
 620  * reads the blocks back in, and checks the crc/md5 of the data.
 621  */
 622 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 623 {
 624         struct fio_file *f;
 625         struct io_u *io_u;
 626         int ret, min_events;
 627         unsigned int i;
 628
 629         dprint(FD_VERIFY, "starting loop\n");
 630
 631         /*
 632          * sync io first and invalidate cache, to make sure we really
 633          * read from disk.
 634          */
 635         for_each_file(td, f, i) {
 636                 if (!fio_file_open(f))
 637                         continue;
 638                 if (fio_io_sync(td, f))
 639                         break;
 640                 if (file_invalidate_cache(td, f))
 641                         break;
 642         }
 643
 644         check_update_rusage(td);
 645
 646         if (td->error)
 647                 return;
 648
 649         /*
 650          * verify_state needs to be reset before verification
 651          * proceeds so that expected random seeds match actual
 652          * random seeds in headers. The main loop will reset
 653          * all random number generators if randrepeat is set.
 654          */
 655         if (!td->o.rand_repeatable)
 656                 td_fill_verify_state_seed(td);
 657
 658         td_set_runstate(td, TD_VERIFYING);
 659
 660         io_u = NULL;
 661         while (!td->terminate) {
 662                 enum fio_ddir ddir;
 663                 int full;
 664
 665                 update_ts_cache(td);
 666                 check_update_rusage(td);
 667
 668                 if (runtime_exceeded(td, &td->ts_cache)) {
 669                         __update_ts_cache(td);
 670                         if (runtime_exceeded(td, &td->ts_cache)) {
 671                                 fio_mark_td_terminate(td);
 672                                 break;
 673                         }
 674                 }
 675
 676                 if (flow_threshold_exceeded(td))
 677                         continue;
 678
 679                 if (!td->o.experimental_verify) {
 680                         io_u = __get_io_u(td);
 681                         if (!io_u)
 682                                 break;
 683
 684                         if (get_next_verify(td, io_u)) {
 685                                 put_io_u(td, io_u);
 686                                 break;
 687                         }
 688
 689                         if (td_io_prep(td, io_u)) {
 690                                 put_io_u(td, io_u);
 691                                 break;
 692                         }
 693                 } else {
 694                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 695                                 break;
 696
 697                         while ((io_u = get_io_u(td)) != NULL) {
 698                                 if (IS_ERR_OR_NULL(io_u)) {
 699                                         io_u = NULL;
 700                                         ret = FIO_Q_BUSY;
 701                                         goto reap;
 702                                 }
 703
 704                                 /*
 705                                  * We are only interested in the places where
 706                                  * we wrote or trimmed IOs. Turn those into
 707                                  * reads for verification purposes.
 708                                  */
 709                                 if (io_u->ddir == DDIR_READ) {
 710                                         /*
 711                                          * Pretend we issued it for rwmix
 712                                          * accounting
 713                                          */
 714                                         td->io_issues[DDIR_READ]++;
 715                                         put_io_u(td, io_u);
 716                                         continue;
 717                                 } else if (io_u->ddir == DDIR_TRIM) {
 718                                         io_u->ddir = DDIR_READ;
 719                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 720                                         break;
 721                                 } else if (io_u->ddir == DDIR_WRITE) {
 722                                         io_u->ddir = DDIR_READ;
 723                                         populate_verify_io_u(td, io_u);
 724                                         break;
 725                                 } else {
 726                                         put_io_u(td, io_u);
 727                                         continue;
 728                                 }
 729                         }
 730
 731                         if (!io_u)
 732                                 break;
 733                 }
 734
 735                 if (verify_state_should_stop(td, io_u)) {
 736                         put_io_u(td, io_u);
 737                         break;
 738                 }
 739
 740                 if (td->o.verify_async)
 741                         io_u->end_io = verify_io_u_async;
 742                 else
 743                         io_u->end_io = verify_io_u;
 744
 745                 ddir = io_u->ddir;
 746                 if (!td->o.disable_slat)
 747                         fio_gettime(&io_u->start_time, NULL);
 748
 749                 ret = io_u_submit(td, io_u);
 750
 751                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 752                         break;
 753
 754                 /*
 755                  * if we can queue more, do so. but check if there are
 756                  * completed io_u's first. Note that we can get BUSY even
 757                  * without IO queued, if the system is resource starved.
 758                  */
 759 reap:
 760                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 761                 if (full || io_in_polling(td))
 762                         ret = wait_for_completions(td, NULL);
 763
 764                 if (ret < 0)
 765                         break;
 766         }
 767
 768         check_update_rusage(td);
 769
 770         if (!td->error) {
 771                 min_events = td->cur_depth;
 772
 773                 if (min_events)
 774                         ret = io_u_queued_complete(td, min_events);
 775         } else
 776                 cleanup_pending_aio(td);
 777
 778         td_set_runstate(td, TD_RUNNING);
 779
 780         dprint(FD_VERIFY, "exiting loop\n");
 781 }
 782
 783 static bool exceeds_number_ios(struct thread_data *td)
 784 {
 785         unsigned long long number_ios;
 786
 787         if (!td->o.number_ios)
 788                 return false;
 789
 790         number_ios = ddir_rw_sum(td->io_blocks);
 791         number_ios += td->io_u_queued + td->io_u_in_flight;
 792
 793         return number_ios >= (td->o.number_ios * td->loops);
 794 }
 795
 796 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 797 {
 798         unsigned long long bytes, limit;
 799
 800         if (td_rw(td))
 801                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 802         else if (td_write(td))
 803                 bytes = this_bytes[DDIR_WRITE];
 804         else if (td_read(td))
 805                 bytes = this_bytes[DDIR_READ];
 806         else
 807                 bytes = this_bytes[DDIR_TRIM];
 808
 809         if (td->o.io_size)
 810                 limit = td->o.io_size;
 811         else
 812                 limit = td->o.size;
 813
 814         limit *= td->loops;
 815         return bytes >= limit || exceeds_number_ios(td);
 816 }
 817
 818 static bool io_issue_bytes_exceeded(struct thread_data *td)
 819 {
 820         return io_bytes_exceeded(td, td->io_issue_bytes);
 821 }
 822
 823 static bool io_complete_bytes_exceeded(struct thread_data *td)
 824 {
 825         return io_bytes_exceeded(td, td->this_io_bytes);
 826 }
 827
 828 /*
 829  * used to calculate the next io time for rate control
 830  *
 831  */
 832 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 833 {
 834         uint64_t bps = td->rate_bps[ddir];
 835
 836         assert(!(td->flags & TD_F_CHILD));
 837
 838         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 839                 uint64_t val, iops;
 840
 841                 iops = bps / td->o.bs[ddir];
 842                 val = (int64_t) (1000000 / iops) *
 843                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 844                 if (val) {
 845                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 846                                         (unsigned long long) 1000000 / val,
 847                                         ddir);
 848                 }
 849                 td->last_usec[ddir] += val;
 850                 return td->last_usec[ddir];
 851         } else if (bps) {
 852                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 853                 uint64_t secs = bytes / bps;
 854                 uint64_t remainder = bytes % bps;
 855
 856                 return remainder * 1000000 / bps + secs * 1000000;
 857         }
 858
 859         return 0;
 860 }
 861
 862 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 863 {
 864         unsigned long long b;
 865         uint64_t total;
 866         int left;
 867
 868         b = ddir_rw_sum(td->io_blocks);
 869         if (b % td->o.thinktime_blocks)
 870                 return;
 871
 872         io_u_quiesce(td);
 873
 874         total = 0;
 875         if (td->o.thinktime_spin)
 876                 total = usec_spin(td->o.thinktime_spin);
 877
 878         left = td->o.thinktime - total;
 879         if (left)
 880                 total += usec_sleep(td, left);
 881
 882         /*
 883          * If we're ignoring thinktime for the rate, add the number of bytes
 884          * we would have done while sleeping, minus one block to ensure we
 885          * start issuing immediately after the sleep.
 886          */
 887         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 888                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 889                 uint64_t bs = td->o.min_bs[ddir];
 890                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 891                 uint64_t over;
 892
 893                 if (usperop <= total)
 894                         over = bs;
 895                 else
 896                         over = (usperop - total) / usperop * -bs;
 897
 898                 td->rate_io_issue_bytes[ddir] += (missed - over);
 899                 /* adjust for rate_process=poisson */
 900                 td->last_usec[ddir] += total;
 901         }
 902 }
 903
 904 /*
 905  * Main IO worker function. It retrieves io_u's to process and queues
 906  * and reaps them, checking for rate and errors along the way.
 907  *
 908  * Returns number of bytes written and trimmed.
 909  */
 910 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 911 {
 912         unsigned int i;
 913         int ret = 0;
 914         uint64_t total_bytes, bytes_issued = 0;
 915
 916         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 917                 bytes_done[i] = td->bytes_done[i];
 918
 919         if (in_ramp_time(td))
 920                 td_set_runstate(td, TD_RAMP);
 921         else
 922                 td_set_runstate(td, TD_RUNNING);
 923
 924         lat_target_init(td);
 925
 926         total_bytes = td->o.size;
 927         /*
 928         * Allow random overwrite workloads to write up to io_size
 929         * before starting verification phase as 'size' doesn't apply.
 930         */
 931         if (td_write(td) && td_random(td) && td->o.norandommap)
 932                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 933         /*
 934          * If verify_backlog is enabled, we'll run the verify in this
 935          * handler as well. For that case, we may need up to twice the
 936          * amount of bytes.
 937          */
 938         if (td->o.verify != VERIFY_NONE &&
 939            (td_write(td) && td->o.verify_backlog))
 940                 total_bytes += td->o.size;
 941
 942         /* In trimwrite mode, each byte is trimmed and then written, so
 943          * allow total_bytes to be twice as big */
 944         if (td_trimwrite(td))
 945                 total_bytes += td->total_io_size;
 946
 947         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 948                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 949                 td->o.time_based) {
 950                 struct timespec comp_time;
 951                 struct io_u *io_u;
 952                 int full;
 953                 enum fio_ddir ddir;
 954
 955                 check_update_rusage(td);
 956
 957                 if (td->terminate || td->done)
 958                         break;
 959
 960                 update_ts_cache(td);
 961
 962                 if (runtime_exceeded(td, &td->ts_cache)) {
 963                         __update_ts_cache(td);
 964                         if (runtime_exceeded(td, &td->ts_cache)) {
 965                                 fio_mark_td_terminate(td);
 966                                 break;
 967                         }
 968                 }
 969
 970                 if (flow_threshold_exceeded(td))
 971                         continue;
 972
 973                 /*
 974                  * Break if we exceeded the bytes. The exception is time
 975                  * based runs, but we still need to break out of the loop
 976                  * for those to run verification, if enabled.
 977                  * Jobs read from iolog do not use this stop condition.
 978                  */
 979                 if (bytes_issued >= total_bytes &&
 980                     !td->o.read_iolog_file &&
 981                     (!td->o.time_based ||
 982                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 983                         break;
 984
 985                 io_u = get_io_u(td);
 986                 if (IS_ERR_OR_NULL(io_u)) {
 987                         int err = PTR_ERR(io_u);
 988
 989                         io_u = NULL;
 990                         ddir = DDIR_INVAL;
 991                         if (err == -EBUSY) {
 992                                 ret = FIO_Q_BUSY;
 993                                 goto reap;
 994                         }
 995                         if (td->o.latency_target)
 996                                 goto reap;
 997                         break;
 998                 }
 999
1000                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1001                         populate_verify_io_u(td, io_u);
1002
1003                 ddir = io_u->ddir;
1004
1005                 /*
1006                  * Add verification end_io handler if:
1007                  *      - Asked to verify (!td_rw(td))
1008                  *      - Or the io_u is from our verify list (mixed write/ver)
1009                  */
1010                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1011                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1012
1013                         if (verify_state_should_stop(td, io_u)) {
1014                                 put_io_u(td, io_u);
1015                                 break;
1016                         }
1017
1018                         if (td->o.verify_async)
1019                                 io_u->end_io = verify_io_u_async;
1020                         else
1021                                 io_u->end_io = verify_io_u;
1022                         td_set_runstate(td, TD_VERIFYING);
1023                 } else if (in_ramp_time(td))
1024                         td_set_runstate(td, TD_RAMP);
1025                 else
1026                         td_set_runstate(td, TD_RUNNING);
1027
1028                 /*
1029                  * Always log IO before it's issued, so we know the specific
1030                  * order of it. The logged unit will track when the IO has
1031                  * completed.
1032                  */
1033                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1034                     td->o.do_verify &&
1035                     td->o.verify != VERIFY_NONE &&
1036                     !td->o.experimental_verify)
1037                         log_io_piece(td, io_u);
1038
1039                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1040                         const unsigned long long blen = io_u->xfer_buflen;
1041                         const enum fio_ddir __ddir = acct_ddir(io_u);
1042
1043                         if (td->error)
1044                                 break;
1045
1046                         workqueue_enqueue(&td->io_wq, &io_u->work);
1047                         ret = FIO_Q_QUEUED;
1048
1049                         if (ddir_rw(__ddir)) {
1050                                 td->io_issues[__ddir]++;
1051                                 td->io_issue_bytes[__ddir] += blen;
1052                                 td->rate_io_issue_bytes[__ddir] += blen;
1053                         }
1054
1055                         if (should_check_rate(td))
1056                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1057
1058                 } else {
1059                         ret = io_u_submit(td, io_u);
1060
1061                         if (should_check_rate(td))
1062                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1063
1064                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1065                                 break;
1066
1067                         /*
1068                          * See if we need to complete some commands. Note that
1069                          * we can get BUSY even without IO queued, if the
1070                          * system is resource starved.
1071                          */
1072 reap:
1073                         full = queue_full(td) ||
1074                                 (ret == FIO_Q_BUSY && td->cur_depth);
1075                         if (full || io_in_polling(td))
1076                                 ret = wait_for_completions(td, &comp_time);
1077                 }
1078                 if (ret < 0)
1079                         break;
1080                 if (!ddir_rw_sum(td->bytes_done) &&
1081                     !td_ioengine_flagged(td, FIO_NOIO))
1082                         continue;
1083
1084                 if (!in_ramp_time(td) && should_check_rate(td)) {
1085                         if (check_min_rate(td, &comp_time)) {
1086                                 if (exitall_on_terminate || td->o.exitall_error)
1087                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1088                                 td_verror(td, EIO, "check_min_rate");
1089                                 break;
1090                         }
1091                 }
1092                 if (!in_ramp_time(td) && td->o.latency_target)
1093                         lat_target_check(td);
1094
1095                 if (ddir_rw(ddir) && td->o.thinktime)
1096                         handle_thinktime(td, ddir);
1097         }
1098
1099         check_update_rusage(td);
1100
1101         if (td->trim_entries)
1102                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1103
1104         if (td->o.fill_device && td->error == ENOSPC) {
1105                 td->error = 0;
1106                 fio_mark_td_terminate(td);
1107         }
1108         if (!td->error) {
1109                 struct fio_file *f;
1110
1111                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1112                         workqueue_flush(&td->io_wq);
1113                         i = 0;
1114                 } else
1115                         i = td->cur_depth;
1116
1117                 if (i) {
1118                         ret = io_u_queued_complete(td, i);
1119                         if (td->o.fill_device && td->error == ENOSPC)
1120                                 td->error = 0;
1121                 }
1122
1123                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1124                         td_set_runstate(td, TD_FSYNCING);
1125
1126                         for_each_file(td, f, i) {
1127                                 if (!fio_file_fsync(td, f))
1128                                         continue;
1129
1130                                 log_err("fio: end_fsync failed for file %s\n",
1131                                                                 f->file_name);
1132                         }
1133                 }
1134         } else
1135                 cleanup_pending_aio(td);
1136
1137         /*
1138          * stop job if we failed doing any IO
1139          */
1140         if (!ddir_rw_sum(td->this_io_bytes))
1141                 td->done = 1;
1142
1143         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1144                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1145 }
1146
1147 static void free_file_completion_logging(struct thread_data *td)
1148 {
1149         struct fio_file *f;
1150         unsigned int i;
1151
1152         for_each_file(td, f, i) {
1153                 if (!f->last_write_comp)
1154                         break;
1155                 sfree(f->last_write_comp);
1156         }
1157 }
1158
1159 static int init_file_completion_logging(struct thread_data *td,
1160                                         unsigned int depth)
1161 {
1162         struct fio_file *f;
1163         unsigned int i;
1164
1165         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1166                 return 0;
1167
1168         for_each_file(td, f, i) {
1169                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1170                 if (!f->last_write_comp)
1171                         goto cleanup;
1172         }
1173
1174         return 0;
1175
1176 cleanup:
1177         free_file_completion_logging(td);
1178         log_err("fio: failed to alloc write comp data\n");
1179         return 1;
1180 }
1181
1182 static void cleanup_io_u(struct thread_data *td)
1183 {
1184         struct io_u *io_u;
1185
1186         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1187
1188                 if (td->io_ops->io_u_free)
1189                         td->io_ops->io_u_free(td, io_u);
1190
1191                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1192         }
1193
1194         free_io_mem(td);
1195
1196         io_u_rexit(&td->io_u_requeues);
1197         io_u_qexit(&td->io_u_freelist, false);
1198         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1199
1200         free_file_completion_logging(td);
1201 }
1202
1203 static int init_io_u(struct thread_data *td)
1204 {
1205         struct io_u *io_u;
1206         int cl_align, i, max_units;
1207         int err;
1208
1209         max_units = td->o.iodepth;
1210
1211         err = 0;
1212         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1213         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1214         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1215
1216         if (err) {
1217                 log_err("fio: failed setting up IO queues\n");
1218                 return 1;
1219         }
1220
1221         cl_align = os_cache_line_size();
1222
1223         for (i = 0; i < max_units; i++) {
1224                 void *ptr;
1225
1226                 if (td->terminate)
1227                         return 1;
1228
1229                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1230                 if (!ptr) {
1231                         log_err("fio: unable to allocate aligned memory\n");
1232                         return 1;
1233                 }
1234
1235                 io_u = ptr;
1236                 memset(io_u, 0, sizeof(*io_u));
1237                 INIT_FLIST_HEAD(&io_u->verify_list);
1238                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1239
1240                 io_u->index = i;
1241                 io_u->flags = IO_U_F_FREE;
1242                 io_u_qpush(&td->io_u_freelist, io_u);
1243
1244                 /*
1245                  * io_u never leaves this stack, used for iteration of all
1246                  * io_u buffers.
1247                  */
1248                 io_u_qpush(&td->io_u_all, io_u);
1249
1250                 if (td->io_ops->io_u_init) {
1251                         int ret = td->io_ops->io_u_init(td, io_u);
1252
1253                         if (ret) {
1254                                 log_err("fio: failed to init engine data: %d\n", ret);
1255                                 return 1;
1256                         }
1257                 }
1258         }
1259
1260         init_io_u_buffers(td);
1261
1262         if (init_file_completion_logging(td, max_units))
1263                 return 1;
1264
1265         return 0;
1266 }
1267
1268 int init_io_u_buffers(struct thread_data *td)
1269 {
1270         struct io_u *io_u;
1271         unsigned long long max_bs, min_write;
1272         int i, max_units;
1273         int data_xfer = 1;
1274         char *p;
1275
1276         max_units = td->o.iodepth;
1277         max_bs = td_max_bs(td);
1278         min_write = td->o.min_bs[DDIR_WRITE];
1279         td->orig_buffer_size = (unsigned long long) max_bs
1280                                         * (unsigned long long) max_units;
1281
1282         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1283                 data_xfer = 0;
1284
1285         /*
1286          * if we may later need to do address alignment, then add any
1287          * possible adjustment here so that we don't cause a buffer
1288          * overflow later. this adjustment may be too much if we get
1289          * lucky and the allocator gives us an aligned address.
1290          */
1291         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1292             td_ioengine_flagged(td, FIO_RAWIO))
1293                 td->orig_buffer_size += page_mask + td->o.mem_align;
1294
1295         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1296                 unsigned long long bs;
1297
1298                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1299                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1300         }
1301
1302         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1303                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1304                 return 1;
1305         }
1306
1307         if (data_xfer && allocate_io_mem(td))
1308                 return 1;
1309
1310         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1311             td_ioengine_flagged(td, FIO_RAWIO))
1312                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1313         else
1314                 p = td->orig_buffer;
1315
1316         for (i = 0; i < max_units; i++) {
1317                 io_u = td->io_u_all.io_us[i];
1318                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1319
1320                 if (data_xfer) {
1321                         io_u->buf = p;
1322                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1323
1324                         if (td_write(td))
1325                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1326                         if (td_write(td) && td->o.verify_pattern_bytes) {
1327                                 /*
1328                                  * Fill the buffer with the pattern if we are
1329                                  * going to be doing writes.
1330                                  */
1331                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1332                         }
1333                 }
1334                 p += max_bs;
1335         }
1336
1337         return 0;
1338 }
1339
1340 /*
1341  * This function is Linux specific.
1342  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1343  */
1344 static int switch_ioscheduler(struct thread_data *td)
1345 {
1346 #ifdef FIO_HAVE_IOSCHED_SWITCH
1347         char tmp[256], tmp2[128], *p;
1348         FILE *f;
1349         int ret;
1350
1351         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1352                 return 0;
1353
1354         assert(td->files && td->files[0]);
1355         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1356
1357         f = fopen(tmp, "r+");
1358         if (!f) {
1359                 if (errno == ENOENT) {
1360                         log_err("fio: os or kernel doesn't support IO scheduler"
1361                                 " switching\n");
1362                         return 0;
1363                 }
1364                 td_verror(td, errno, "fopen iosched");
1365                 return 1;
1366         }
1367
1368         /*
1369          * Set io scheduler.
1370          */
1371         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1372         if (ferror(f) || ret != 1) {
1373                 td_verror(td, errno, "fwrite");
1374                 fclose(f);
1375                 return 1;
1376         }
1377
1378         rewind(f);
1379
1380         /*
1381          * Read back and check that the selected scheduler is now the default.
1382          */
1383         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1384         if (ferror(f) || ret < 0) {
1385                 td_verror(td, errno, "fread");
1386                 fclose(f);
1387                 return 1;
1388         }
1389         tmp[ret] = '\0';
1390         /*
1391          * either a list of io schedulers or "none\n" is expected. Strip the
1392          * trailing newline.
1393          */
1394         p = tmp;
1395         strsep(&p, "\n");
1396
1397         /*
1398          * Write to "none" entry doesn't fail, so check the result here.
1399          */
1400         if (!strcmp(tmp, "none")) {
1401                 log_err("fio: io scheduler is not tunable\n");
1402                 fclose(f);
1403                 return 0;
1404         }
1405
1406         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1407         if (!strstr(tmp, tmp2)) {
1408                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1409                 td_verror(td, EINVAL, "iosched_switch");
1410                 fclose(f);
1411                 return 1;
1412         }
1413
1414         fclose(f);
1415         return 0;
1416 #else
1417         return 0;
1418 #endif
1419 }
1420
1421 static bool keep_running(struct thread_data *td)
1422 {
1423         unsigned long long limit;
1424
1425         if (td->done)
1426                 return false;
1427         if (td->terminate)
1428                 return false;
1429         if (td->o.time_based)
1430                 return true;
1431         if (td->o.loops) {
1432                 td->o.loops--;
1433                 return true;
1434         }
1435         if (exceeds_number_ios(td))
1436                 return false;
1437
1438         if (td->o.io_size)
1439                 limit = td->o.io_size;
1440         else
1441                 limit = td->o.size;
1442
1443         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1444                 uint64_t diff;
1445
1446                 /*
1447                  * If the difference is less than the maximum IO size, we
1448                  * are done.
1449                  */
1450                 diff = limit - ddir_rw_sum(td->io_bytes);
1451                 if (diff < td_max_bs(td))
1452                         return false;
1453
1454                 if (fio_files_done(td) && !td->o.io_size)
1455                         return false;
1456
1457                 return true;
1458         }
1459
1460         return false;
1461 }
1462
1463 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1464 {
1465         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 13 + 1;
1466         int ret;
1467         char *str;
1468
1469         str = malloc(newlen);
1470         sprintf(str, "%s > %s.%s.txt 2>&1", string, o->name, mode);
1471
1472         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1473         ret = system(str);
1474         if (ret == -1)
1475                 log_err("fio: exec of cmd <%s> failed\n", str);
1476
1477         free(str);
1478         return ret;
1479 }
1480
1481 /*
1482  * Dry run to compute correct state of numberio for verification.
1483  */
1484 static uint64_t do_dry_run(struct thread_data *td)
1485 {
1486         td_set_runstate(td, TD_RUNNING);
1487
1488         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1489                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1490                 struct io_u *io_u;
1491                 int ret;
1492
1493                 if (td->terminate || td->done)
1494                         break;
1495
1496                 io_u = get_io_u(td);
1497                 if (IS_ERR_OR_NULL(io_u))
1498                         break;
1499
1500                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1501                 io_u->error = 0;
1502                 io_u->resid = 0;
1503                 if (ddir_rw(acct_ddir(io_u)))
1504                         td->io_issues[acct_ddir(io_u)]++;
1505                 if (ddir_rw(io_u->ddir)) {
1506                         io_u_mark_depth(td, 1);
1507                         td->ts.total_io_u[io_u->ddir]++;
1508                 }
1509
1510                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1511                     td->o.do_verify &&
1512                     td->o.verify != VERIFY_NONE &&
1513                     !td->o.experimental_verify)
1514                         log_io_piece(td, io_u);
1515
1516                 ret = io_u_sync_complete(td, io_u);
1517                 (void) ret;
1518         }
1519
1520         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1521 }
1522
1523 struct fork_data {
1524         struct thread_data *td;
1525         struct sk_out *sk_out;
1526 };
1527
1528 /*
1529  * Entry point for the thread based jobs. The process based jobs end up
1530  * here as well, after a little setup.
1531  */
1532 static void *thread_main(void *data)
1533 {
1534         struct fork_data *fd = data;
1535         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1536         struct thread_data *td = fd->td;
1537         struct thread_options *o = &td->o;
1538         struct sk_out *sk_out = fd->sk_out;
1539         uint64_t bytes_done[DDIR_RWDIR_CNT];
1540         int deadlock_loop_cnt;
1541         bool clear_state;
1542         int res, ret;
1543
1544         sk_out_assign(sk_out);
1545         free(fd);
1546
1547         if (!o->use_thread) {
1548                 setsid();
1549                 td->pid = getpid();
1550         } else
1551                 td->pid = gettid();
1552
1553         fio_local_clock_init();
1554
1555         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1556
1557         if (is_backend)
1558                 fio_server_send_start(td);
1559
1560         INIT_FLIST_HEAD(&td->io_log_list);
1561         INIT_FLIST_HEAD(&td->io_hist_list);
1562         INIT_FLIST_HEAD(&td->verify_list);
1563         INIT_FLIST_HEAD(&td->trim_list);
1564         td->io_hist_tree = RB_ROOT;
1565
1566         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1567         if (ret) {
1568                 td_verror(td, ret, "mutex_cond_init_pshared");
1569                 goto err;
1570         }
1571         ret = cond_init_pshared(&td->verify_cond);
1572         if (ret) {
1573                 td_verror(td, ret, "mutex_cond_pshared");
1574                 goto err;
1575         }
1576
1577         td_set_runstate(td, TD_INITIALIZED);
1578         dprint(FD_MUTEX, "up startup_sem\n");
1579         fio_sem_up(startup_sem);
1580         dprint(FD_MUTEX, "wait on td->sem\n");
1581         fio_sem_down(td->sem);
1582         dprint(FD_MUTEX, "done waiting on td->sem\n");
1583
1584         /*
1585          * A new gid requires privilege, so we need to do this before setting
1586          * the uid.
1587          */
1588         if (o->gid != -1U && setgid(o->gid)) {
1589                 td_verror(td, errno, "setgid");
1590                 goto err;
1591         }
1592         if (o->uid != -1U && setuid(o->uid)) {
1593                 td_verror(td, errno, "setuid");
1594                 goto err;
1595         }
1596
1597         td_zone_gen_index(td);
1598
1599         /*
1600          * Do this early, we don't want the compress threads to be limited
1601          * to the same CPUs as the IO workers. So do this before we set
1602          * any potential CPU affinity
1603          */
1604         if (iolog_compress_init(td, sk_out))
1605                 goto err;
1606
1607         /*
1608          * If we have a gettimeofday() thread, make sure we exclude that
1609          * thread from this job
1610          */
1611         if (o->gtod_cpu)
1612                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1613
1614         /*
1615          * Set affinity first, in case it has an impact on the memory
1616          * allocations.
1617          */
1618         if (fio_option_is_set(o, cpumask)) {
1619                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1620                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1621                         if (!ret) {
1622                                 log_err("fio: no CPUs set\n");
1623                                 log_err("fio: Try increasing number of available CPUs\n");
1624                                 td_verror(td, EINVAL, "cpus_split");
1625                                 goto err;
1626                         }
1627                 }
1628                 ret = fio_setaffinity(td->pid, o->cpumask);
1629                 if (ret == -1) {
1630                         td_verror(td, errno, "cpu_set_affinity");
1631                         goto err;
1632                 }
1633         }
1634
1635 #ifdef CONFIG_LIBNUMA
1636         /* numa node setup */
1637         if (fio_option_is_set(o, numa_cpunodes) ||
1638             fio_option_is_set(o, numa_memnodes)) {
1639                 struct bitmask *mask;
1640
1641                 if (numa_available() < 0) {
1642                         td_verror(td, errno, "Does not support NUMA API\n");
1643                         goto err;
1644                 }
1645
1646                 if (fio_option_is_set(o, numa_cpunodes)) {
1647                         mask = numa_parse_nodestring(o->numa_cpunodes);
1648                         ret = numa_run_on_node_mask(mask);
1649                         numa_free_nodemask(mask);
1650                         if (ret == -1) {
1651                                 td_verror(td, errno, \
1652                                         "numa_run_on_node_mask failed\n");
1653                                 goto err;
1654                         }
1655                 }
1656
1657                 if (fio_option_is_set(o, numa_memnodes)) {
1658                         mask = NULL;
1659                         if (o->numa_memnodes)
1660                                 mask = numa_parse_nodestring(o->numa_memnodes);
1661
1662                         switch (o->numa_mem_mode) {
1663                         case MPOL_INTERLEAVE:
1664                                 numa_set_interleave_mask(mask);
1665                                 break;
1666                         case MPOL_BIND:
1667                                 numa_set_membind(mask);
1668                                 break;
1669                         case MPOL_LOCAL:
1670                                 numa_set_localalloc();
1671                                 break;
1672                         case MPOL_PREFERRED:
1673                                 numa_set_preferred(o->numa_mem_prefer_node);
1674                                 break;
1675                         case MPOL_DEFAULT:
1676                         default:
1677                                 break;
1678                         }
1679
1680                         if (mask)
1681                                 numa_free_nodemask(mask);
1682
1683                 }
1684         }
1685 #endif
1686
1687         if (fio_pin_memory(td))
1688                 goto err;
1689
1690         /*
1691          * May alter parameters that init_io_u() will use, so we need to
1692          * do this first.
1693          */
1694         if (!init_iolog(td))
1695                 goto err;
1696
1697         if (td_io_init(td))
1698                 goto err;
1699
1700         if (init_io_u(td))
1701                 goto err;
1702
1703         if (td->io_ops->post_init && td->io_ops->post_init(td))
1704                 goto err;
1705
1706         if (o->verify_async && verify_async_init(td))
1707                 goto err;
1708
1709         if (fio_option_is_set(o, ioprio) ||
1710             fio_option_is_set(o, ioprio_class)) {
1711                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1712                 if (ret == -1) {
1713                         td_verror(td, errno, "ioprio_set");
1714                         goto err;
1715                 }
1716         }
1717
1718         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1719                 goto err;
1720
1721         errno = 0;
1722         if (nice(o->nice) == -1 && errno != 0) {
1723                 td_verror(td, errno, "nice");
1724                 goto err;
1725         }
1726
1727         if (o->ioscheduler && switch_ioscheduler(td))
1728                 goto err;
1729
1730         if (!o->create_serialize && setup_files(td))
1731                 goto err;
1732
1733         if (!init_random_map(td))
1734                 goto err;
1735
1736         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1737                 goto err;
1738
1739         if (o->pre_read && !pre_read_files(td))
1740                 goto err;
1741
1742         fio_verify_init(td);
1743
1744         if (rate_submit_init(td, sk_out))
1745                 goto err;
1746
1747         set_epoch_time(td, o->log_unix_epoch);
1748         fio_getrusage(&td->ru_start);
1749         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1750         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1751         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1752
1753         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1754                         o->ratemin[DDIR_TRIM]) {
1755                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1756                                         sizeof(td->bw_sample_time));
1757                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1758                                         sizeof(td->bw_sample_time));
1759                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1760                                         sizeof(td->bw_sample_time));
1761         }
1762
1763         memset(bytes_done, 0, sizeof(bytes_done));
1764         clear_state = false;
1765
1766         while (keep_running(td)) {
1767                 uint64_t verify_bytes;
1768
1769                 fio_gettime(&td->start, NULL);
1770                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1771
1772                 if (clear_state) {
1773                         clear_io_state(td, 0);
1774
1775                         if (o->unlink_each_loop && unlink_all_files(td))
1776                                 break;
1777                 }
1778
1779                 prune_io_piece_log(td);
1780
1781                 if (td->o.verify_only && td_write(td))
1782                         verify_bytes = do_dry_run(td);
1783                 else {
1784                         do_io(td, bytes_done);
1785
1786                         if (!ddir_rw_sum(bytes_done)) {
1787                                 fio_mark_td_terminate(td);
1788                                 verify_bytes = 0;
1789                         } else {
1790                                 verify_bytes = bytes_done[DDIR_WRITE] +
1791                                                 bytes_done[DDIR_TRIM];
1792                         }
1793                 }
1794
1795                 /*
1796                  * If we took too long to shut down, the main thread could
1797                  * already consider us reaped/exited. If that happens, break
1798                  * out and clean up.
1799                  */
1800                 if (td->runstate >= TD_EXITED)
1801                         break;
1802
1803                 clear_state = true;
1804
1805                 /*
1806                  * Make sure we've successfully updated the rusage stats
1807                  * before waiting on the stat mutex. Otherwise we could have
1808                  * the stat thread holding stat mutex and waiting for
1809                  * the rusage_sem, which would never get upped because
1810                  * this thread is waiting for the stat mutex.
1811                  */
1812                 deadlock_loop_cnt = 0;
1813                 do {
1814                         check_update_rusage(td);
1815                         if (!fio_sem_down_trylock(stat_sem))
1816                                 break;
1817                         usleep(1000);
1818                         if (deadlock_loop_cnt++ > 5000) {
1819                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1820                                 td->error = EDEADLK;
1821                                 goto err;
1822                         }
1823                 } while (1);
1824
1825                 if (td_read(td) && td->io_bytes[DDIR_READ])
1826                         update_runtime(td, elapsed_us, DDIR_READ);
1827                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1828                         update_runtime(td, elapsed_us, DDIR_WRITE);
1829                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1830                         update_runtime(td, elapsed_us, DDIR_TRIM);
1831                 fio_gettime(&td->start, NULL);
1832                 fio_sem_up(stat_sem);
1833
1834                 if (td->error || td->terminate)
1835                         break;
1836
1837                 if (!o->do_verify ||
1838                     o->verify == VERIFY_NONE ||
1839                     td_ioengine_flagged(td, FIO_UNIDIR))
1840                         continue;
1841
1842                 clear_io_state(td, 0);
1843
1844                 fio_gettime(&td->start, NULL);
1845
1846                 do_verify(td, verify_bytes);
1847
1848                 /*
1849                  * See comment further up for why this is done here.
1850                  */
1851                 check_update_rusage(td);
1852
1853                 fio_sem_down(stat_sem);
1854                 update_runtime(td, elapsed_us, DDIR_READ);
1855                 fio_gettime(&td->start, NULL);
1856                 fio_sem_up(stat_sem);
1857
1858                 if (td->error || td->terminate)
1859                         break;
1860         }
1861
1862         /*
1863          * Acquire this lock if we were doing overlap checking in
1864          * offload mode so that we don't clean up this job while
1865          * another thread is checking its io_u's for overlap
1866          */
1867         if (td_offload_overlap(td)) {
1868                 int res = pthread_mutex_lock(&overlap_check);
1869                 assert(res == 0);
1870         }
1871         td_set_runstate(td, TD_FINISHING);
1872         if (td_offload_overlap(td)) {
1873                 res = pthread_mutex_unlock(&overlap_check);
1874                 assert(res == 0);
1875         }
1876
1877         update_rusage_stat(td);
1878         td->ts.total_run_time = mtime_since_now(&td->epoch);
1879         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1880         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1881         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1882
1883         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1884             (td->o.verify != VERIFY_NONE && td_write(td)))
1885                 verify_save_state(td->thread_number);
1886
1887         fio_unpin_memory(td);
1888
1889         td_writeout_logs(td, true);
1890
1891         iolog_compress_exit(td);
1892         rate_submit_exit(td);
1893
1894         if (o->exec_postrun)
1895                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1896
1897         if (exitall_on_terminate || (o->exitall_error && td->error))
1898                 fio_terminate_threads(td->groupid, td->o.exit_what);
1899
1900 err:
1901         if (td->error)
1902                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1903                                                         td->verror);
1904
1905         if (o->verify_async)
1906                 verify_async_exit(td);
1907
1908         close_and_free_files(td);
1909         cleanup_io_u(td);
1910         close_ioengine(td);
1911         cgroup_shutdown(td, cgroup_mnt);
1912         verify_free_state(td);
1913         td_zone_free_index(td);
1914
1915         if (fio_option_is_set(o, cpumask)) {
1916                 ret = fio_cpuset_exit(&o->cpumask);
1917                 if (ret)
1918                         td_verror(td, ret, "fio_cpuset_exit");
1919         }
1920
1921         /*
1922          * do this very late, it will log file closing as well
1923          */
1924         if (o->write_iolog_file)
1925                 write_iolog_close(td);
1926         if (td->io_log_rfile)
1927                 fclose(td->io_log_rfile);
1928
1929         td_set_runstate(td, TD_EXITED);
1930
1931         /*
1932          * Do this last after setting our runstate to exited, so we
1933          * know that the stat thread is signaled.
1934          */
1935         check_update_rusage(td);
1936
1937         sk_out_drop();
1938         return (void *) (uintptr_t) td->error;
1939 }
1940
1941 /*
1942  * Run over the job map and reap the threads that have exited, if any.
1943  */
1944 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1945                          uint64_t *m_rate)
1946 {
1947         struct thread_data *td;
1948         unsigned int cputhreads, realthreads, pending;
1949         int i, status, ret;
1950
1951         /*
1952          * reap exited threads (TD_EXITED -> TD_REAPED)
1953          */
1954         realthreads = pending = cputhreads = 0;
1955         for_each_td(td, i) {
1956                 int flags = 0;
1957
1958                  if (!strcmp(td->o.ioengine, "cpuio"))
1959                         cputhreads++;
1960                 else
1961                         realthreads++;
1962
1963                 if (!td->pid) {
1964                         pending++;
1965                         continue;
1966                 }
1967                 if (td->runstate == TD_REAPED)
1968                         continue;
1969                 if (td->o.use_thread) {
1970                         if (td->runstate == TD_EXITED) {
1971                                 td_set_runstate(td, TD_REAPED);
1972                                 goto reaped;
1973                         }
1974                         continue;
1975                 }
1976
1977                 flags = WNOHANG;
1978                 if (td->runstate == TD_EXITED)
1979                         flags = 0;
1980
1981                 /*
1982                  * check if someone quit or got killed in an unusual way
1983                  */
1984                 ret = waitpid(td->pid, &status, flags);
1985                 if (ret < 0) {
1986                         if (errno == ECHILD) {
1987                                 log_err("fio: pid=%d disappeared %d\n",
1988                                                 (int) td->pid, td->runstate);
1989                                 td->sig = ECHILD;
1990                                 td_set_runstate(td, TD_REAPED);
1991                                 goto reaped;
1992                         }
1993                         perror("waitpid");
1994                 } else if (ret == td->pid) {
1995                         if (WIFSIGNALED(status)) {
1996                                 int sig = WTERMSIG(status);
1997
1998                                 if (sig != SIGTERM && sig != SIGUSR2)
1999                                         log_err("fio: pid=%d, got signal=%d\n",
2000                                                         (int) td->pid, sig);
2001                                 td->sig = sig;
2002                                 td_set_runstate(td, TD_REAPED);
2003                                 goto reaped;
2004                         }
2005                         if (WIFEXITED(status)) {
2006                                 if (WEXITSTATUS(status) && !td->error)
2007                                         td->error = WEXITSTATUS(status);
2008
2009                                 td_set_runstate(td, TD_REAPED);
2010                                 goto reaped;
2011                         }
2012                 }
2013
2014                 /*
2015                  * If the job is stuck, do a forceful timeout of it and
2016                  * move on.
2017                  */
2018                 if (td->terminate &&
2019                     td->runstate < TD_FSYNCING &&
2020                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2021                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2022                                 "%lu seconds, it appears to be stuck. Doing "
2023                                 "forceful exit of this job.\n",
2024                                 td->o.name, td->runstate,
2025                                 (unsigned long) time_since_now(&td->terminate_time));
2026                         td_set_runstate(td, TD_REAPED);
2027                         goto reaped;
2028                 }
2029
2030                 /*
2031                  * thread is not dead, continue
2032                  */
2033                 pending++;
2034                 continue;
2035 reaped:
2036                 (*nr_running)--;
2037                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2038                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2039                 if (!td->pid)
2040                         pending--;
2041
2042                 if (td->error)
2043                         exit_value++;
2044
2045                 done_secs += mtime_since_now(&td->epoch) / 1000;
2046                 profile_td_exit(td);
2047         }
2048
2049         if (*nr_running == cputhreads && !pending && realthreads)
2050                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2051 }
2052
2053 static bool __check_trigger_file(void)
2054 {
2055         struct stat sb;
2056
2057         if (!trigger_file)
2058                 return false;
2059
2060         if (stat(trigger_file, &sb))
2061                 return false;
2062
2063         if (unlink(trigger_file) < 0)
2064                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2065                                                         strerror(errno));
2066
2067         return true;
2068 }
2069
2070 static bool trigger_timedout(void)
2071 {
2072         if (trigger_timeout)
2073                 if (time_since_genesis() >= trigger_timeout) {
2074                         trigger_timeout = 0;
2075                         return true;
2076                 }
2077
2078         return false;
2079 }
2080
2081 void exec_trigger(const char *cmd)
2082 {
2083         int ret;
2084
2085         if (!cmd || cmd[0] == '\0')
2086                 return;
2087
2088         ret = system(cmd);
2089         if (ret == -1)
2090                 log_err("fio: failed executing %s trigger\n", cmd);
2091 }
2092
2093 void check_trigger_file(void)
2094 {
2095         if (__check_trigger_file() || trigger_timedout()) {
2096                 if (nr_clients)
2097                         fio_clients_send_trigger(trigger_remote_cmd);
2098                 else {
2099                         verify_save_state(IO_LIST_ALL);
2100                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2101                         exec_trigger(trigger_cmd);
2102                 }
2103         }
2104 }
2105
2106 static int fio_verify_load_state(struct thread_data *td)
2107 {
2108         int ret;
2109
2110         if (!td->o.verify_state)
2111                 return 0;
2112
2113         if (is_backend) {
2114                 void *data;
2115
2116                 ret = fio_server_get_verify_state(td->o.name,
2117                                         td->thread_number - 1, &data);
2118                 if (!ret)
2119                         verify_assign_state(td, data);
2120         } else {
2121                 char prefix[PATH_MAX];
2122
2123                 if (aux_path)
2124                         sprintf(prefix, "%s%clocal", aux_path,
2125                                         FIO_OS_PATH_SEPARATOR);
2126                 else
2127                         strcpy(prefix, "local");
2128                 ret = verify_load_state(td, prefix);
2129         }
2130
2131         return ret;
2132 }
2133
2134 static void do_usleep(unsigned int usecs)
2135 {
2136         check_for_running_stats();
2137         check_trigger_file();
2138         usleep(usecs);
2139 }
2140
2141 static bool check_mount_writes(struct thread_data *td)
2142 {
2143         struct fio_file *f;
2144         unsigned int i;
2145
2146         if (!td_write(td) || td->o.allow_mounted_write)
2147                 return false;
2148
2149         /*
2150          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2151          * are mkfs'd and mounted.
2152          */
2153         for_each_file(td, f, i) {
2154 #ifdef FIO_HAVE_CHARDEV_SIZE
2155                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2156 #else
2157                 if (f->filetype != FIO_TYPE_BLOCK)
2158 #endif
2159                         continue;
2160                 if (device_is_mounted(f->file_name))
2161                         goto mounted;
2162         }
2163
2164         return false;
2165 mounted:
2166         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2167         return true;
2168 }
2169
2170 static bool waitee_running(struct thread_data *me)
2171 {
2172         const char *waitee = me->o.wait_for;
2173         const char *self = me->o.name;
2174         struct thread_data *td;
2175         int i;
2176
2177         if (!waitee)
2178                 return false;
2179
2180         for_each_td(td, i) {
2181                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2182                         continue;
2183
2184                 if (td->runstate < TD_EXITED) {
2185                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2186                                         self, td->o.name,
2187                                         runstate_to_name(td->runstate));
2188                         return true;
2189                 }
2190         }
2191
2192         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2193         return false;
2194 }
2195
2196 /*
2197  * Main function for kicking off and reaping jobs, as needed.
2198  */
2199 static void run_threads(struct sk_out *sk_out)
2200 {
2201         struct thread_data *td;
2202         unsigned int i, todo, nr_running, nr_started;
2203         uint64_t m_rate, t_rate;
2204         uint64_t spent;
2205
2206         if (fio_gtod_offload && fio_start_gtod_thread())
2207                 return;
2208
2209         fio_idle_prof_init();
2210
2211         set_sig_handlers();
2212
2213         nr_thread = nr_process = 0;
2214         for_each_td(td, i) {
2215                 if (check_mount_writes(td))
2216                         return;
2217                 if (td->o.use_thread)
2218                         nr_thread++;
2219                 else
2220                         nr_process++;
2221         }
2222
2223         if (output_format & FIO_OUTPUT_NORMAL) {
2224                 struct buf_output out;
2225
2226                 buf_output_init(&out);
2227                 __log_buf(&out, "Starting ");
2228                 if (nr_thread)
2229                         __log_buf(&out, "%d thread%s", nr_thread,
2230                                                 nr_thread > 1 ? "s" : "");
2231                 if (nr_process) {
2232                         if (nr_thread)
2233                                 __log_buf(&out, " and ");
2234                         __log_buf(&out, "%d process%s", nr_process,
2235                                                 nr_process > 1 ? "es" : "");
2236                 }
2237                 __log_buf(&out, "\n");
2238                 log_info_buf(out.buf, out.buflen);
2239                 buf_output_free(&out);
2240         }
2241
2242         todo = thread_number;
2243         nr_running = 0;
2244         nr_started = 0;
2245         m_rate = t_rate = 0;
2246
2247         for_each_td(td, i) {
2248                 print_status_init(td->thread_number - 1);
2249
2250                 if (!td->o.create_serialize)
2251                         continue;
2252
2253                 if (fio_verify_load_state(td))
2254                         goto reap;
2255
2256                 /*
2257                  * do file setup here so it happens sequentially,
2258                  * we don't want X number of threads getting their
2259                  * client data interspersed on disk
2260                  */
2261                 if (setup_files(td)) {
2262 reap:
2263                         exit_value++;
2264                         if (td->error)
2265                                 log_err("fio: pid=%d, err=%d/%s\n",
2266                                         (int) td->pid, td->error, td->verror);
2267                         td_set_runstate(td, TD_REAPED);
2268                         todo--;
2269                 } else {
2270                         struct fio_file *f;
2271                         unsigned int j;
2272
2273                         /*
2274                          * for sharing to work, each job must always open
2275                          * its own files. so close them, if we opened them
2276                          * for creation
2277                          */
2278                         for_each_file(td, f, j) {
2279                                 if (fio_file_open(f))
2280                                         td_io_close_file(td, f);
2281                         }
2282                 }
2283         }
2284
2285         /* start idle threads before io threads start to run */
2286         fio_idle_prof_start();
2287
2288         set_genesis_time();
2289
2290         while (todo) {
2291                 struct thread_data *map[REAL_MAX_JOBS];
2292                 struct timespec this_start;
2293                 int this_jobs = 0, left;
2294                 struct fork_data *fd;
2295
2296                 /*
2297                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2298                  */
2299                 for_each_td(td, i) {
2300                         if (td->runstate != TD_NOT_CREATED)
2301                                 continue;
2302
2303                         /*
2304                          * never got a chance to start, killed by other
2305                          * thread for some reason
2306                          */
2307                         if (td->terminate) {
2308                                 todo--;
2309                                 continue;
2310                         }
2311
2312                         if (td->o.start_delay) {
2313                                 spent = utime_since_genesis();
2314
2315                                 if (td->o.start_delay > spent)
2316                                         continue;
2317                         }
2318
2319                         if (td->o.stonewall && (nr_started || nr_running)) {
2320                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2321                                                         td->o.name);
2322                                 break;
2323                         }
2324
2325                         if (waitee_running(td)) {
2326                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2327                                                 td->o.name, td->o.wait_for);
2328                                 continue;
2329                         }
2330
2331                         init_disk_util(td);
2332
2333                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2334                         td->update_rusage = 0;
2335
2336                         /*
2337                          * Set state to created. Thread will transition
2338                          * to TD_INITIALIZED when it's done setting up.
2339                          */
2340                         td_set_runstate(td, TD_CREATED);
2341                         map[this_jobs++] = td;
2342                         nr_started++;
2343
2344                         fd = calloc(1, sizeof(*fd));
2345                         fd->td = td;
2346                         fd->sk_out = sk_out;
2347
2348                         if (td->o.use_thread) {
2349                                 int ret;
2350
2351                                 dprint(FD_PROCESS, "will pthread_create\n");
2352                                 ret = pthread_create(&td->thread, NULL,
2353                                                         thread_main, fd);
2354                                 if (ret) {
2355                                         log_err("pthread_create: %s\n",
2356                                                         strerror(ret));
2357                                         free(fd);
2358                                         nr_started--;
2359                                         break;
2360                                 }
2361                                 fd = NULL;
2362                                 ret = pthread_detach(td->thread);
2363                                 if (ret)
2364                                         log_err("pthread_detach: %s",
2365                                                         strerror(ret));
2366                         } else {
2367                                 pid_t pid;
2368                                 dprint(FD_PROCESS, "will fork\n");
2369                                 pid = fork();
2370                                 if (!pid) {
2371                                         int ret;
2372
2373                                         ret = (int)(uintptr_t)thread_main(fd);
2374                                         _exit(ret);
2375                                 } else if (i == fio_debug_jobno)
2376                                         *fio_debug_jobp = pid;
2377                         }
2378                         dprint(FD_MUTEX, "wait on startup_sem\n");
2379                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2380                                 log_err("fio: job startup hung? exiting.\n");
2381                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2382                                 fio_abort = true;
2383                                 nr_started--;
2384                                 free(fd);
2385                                 break;
2386                         }
2387                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2388                 }
2389
2390                 /*
2391                  * Wait for the started threads to transition to
2392                  * TD_INITIALIZED.
2393                  */
2394                 fio_gettime(&this_start, NULL);
2395                 left = this_jobs;
2396                 while (left && !fio_abort) {
2397                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2398                                 break;
2399
2400                         do_usleep(100000);
2401
2402                         for (i = 0; i < this_jobs; i++) {
2403                                 td = map[i];
2404                                 if (!td)
2405                                         continue;
2406                                 if (td->runstate == TD_INITIALIZED) {
2407                                         map[i] = NULL;
2408                                         left--;
2409                                 } else if (td->runstate >= TD_EXITED) {
2410                                         map[i] = NULL;
2411                                         left--;
2412                                         todo--;
2413                                         nr_running++; /* work-around... */
2414                                 }
2415                         }
2416                 }
2417
2418                 if (left) {
2419                         log_err("fio: %d job%s failed to start\n", left,
2420                                         left > 1 ? "s" : "");
2421                         for (i = 0; i < this_jobs; i++) {
2422                                 td = map[i];
2423                                 if (!td)
2424                                         continue;
2425                                 kill(td->pid, SIGTERM);
2426                         }
2427                         break;
2428                 }
2429
2430                 /*
2431                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2432                  */
2433                 for_each_td(td, i) {
2434                         if (td->runstate != TD_INITIALIZED)
2435                                 continue;
2436
2437                         if (in_ramp_time(td))
2438                                 td_set_runstate(td, TD_RAMP);
2439                         else
2440                                 td_set_runstate(td, TD_RUNNING);
2441                         nr_running++;
2442                         nr_started--;
2443                         m_rate += ddir_rw_sum(td->o.ratemin);
2444                         t_rate += ddir_rw_sum(td->o.rate);
2445                         todo--;
2446                         fio_sem_up(td->sem);
2447                 }
2448
2449                 reap_threads(&nr_running, &t_rate, &m_rate);
2450
2451                 if (todo)
2452                         do_usleep(100000);
2453         }
2454
2455         while (nr_running) {
2456                 reap_threads(&nr_running, &t_rate, &m_rate);
2457                 do_usleep(10000);
2458         }
2459
2460         fio_idle_prof_stop();
2461
2462         update_io_ticks();
2463 }
2464
2465 static void free_disk_util(void)
2466 {
2467         disk_util_prune_entries();
2468         helper_thread_destroy();
2469 }
2470
2471 int fio_backend(struct sk_out *sk_out)
2472 {
2473         struct thread_data *td;
2474         int i;
2475
2476         if (exec_profile) {
2477                 if (load_profile(exec_profile))
2478                         return 1;
2479                 free(exec_profile);
2480                 exec_profile = NULL;
2481         }
2482         if (!thread_number)
2483                 return 0;
2484
2485         if (write_bw_log) {
2486                 struct log_params p = {
2487                         .log_type = IO_LOG_TYPE_BW,
2488                 };
2489
2490                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2491                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2492                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2493         }
2494
2495         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2496         if (!sk_out)
2497                 is_local_backend = true;
2498         if (startup_sem == NULL)
2499                 return 1;
2500
2501         set_genesis_time();
2502         stat_init();
2503         if (helper_thread_create(startup_sem, sk_out))
2504                 log_err("fio: failed to create helper thread\n");
2505
2506         cgroup_list = smalloc(sizeof(*cgroup_list));
2507         if (cgroup_list)
2508                 INIT_FLIST_HEAD(cgroup_list);
2509
2510         run_threads(sk_out);
2511
2512         helper_thread_exit();
2513
2514         if (!fio_abort) {
2515                 __show_run_stats();
2516                 if (write_bw_log) {
2517                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2518                                 struct io_log *log = agg_io_log[i];
2519
2520                                 flush_log(log, false);
2521                                 free_log(log);
2522                         }
2523                 }
2524         }
2525
2526         for_each_td(td, i) {
2527                 steadystate_free(td);
2528                 fio_options_free(td);
2529                 if (td->rusage_sem) {
2530                         fio_sem_remove(td->rusage_sem);
2531                         td->rusage_sem = NULL;
2532                 }
2533                 fio_sem_remove(td->sem);
2534                 td->sem = NULL;
2535         }
2536
2537         free_disk_util();
2538         if (cgroup_list) {
2539                 cgroup_kill(cgroup_list);
2540                 sfree(cgroup_list);
2541         }
2542
2543         fio_sem_remove(startup_sem);
2544         stat_exit();
2545         return exit_value;
2546 }